Repository navigation
feat: add ai mcp stdio MCP server - #110
Merged
Merged
Conversation
`ai mcp [--profile core|media|all]` serves newline-delimited JSON-RPC (MCP 2025-11-25, tools only). Each tool call runs the same `ai` binary with the typed arguments turned into argv, so the CLI stays the only place that does the work. Tool schemas come from the registered command tree through a new read-only `Command.definition()` accessor. A small override table adds what the parser cannot know: profile, forced --json, required output on video start, hidden plumbing flags, and the typed ai_decide questions. Child runs ignore stdin (except decide state), have no controlling TTY, run with color off, and are killed as a process group on timeout, stdout close, or SIGTERM/SIGINT/SIGHUP.
Unit tests for the argv translation, validation messages, and result mapping, plus one end-to-end stdio session against the built binary that compares ai_models with plain `ai models --json`. A sync test walks the CLI command tree and fails when a command is neither a tool nor explicitly excluded, or when a command option is neither exposed nor hidden, so CLI changes cannot silently drift from the MCP tools.
Add an MCP page to the website docs and nav, an MCP section to the README, the command to the agent skill and footer, and an Unreleased changelog entry. The release markers stay on 0.7.0. AGENTS.md now says how to keep MCP in step with the CLI: map or exclude new commands and hide flags agents should not see; the sync test enforces it.
Drop the duplicated question schema, unreachable checks, and the second input pass; move decide's replaced flags into its override; skip blank stdin lines. Require text and audio on the speech tools, since MCP runs have no stdin to fall back on. Tests now cover only behavior that can break: argv encoding, input the CLI would misread, result mapping, one stdio session against the built binary, and the CLI sync check. The sync check walks every runnable command (new `runnable` field on Command.definition()), so a command with subcommands and its own action, like `video`, is not missed.
opencode prefixes tool ids with the server key, so a key of `ai` shows tools as `ai_ai_text`. Found while dogfooding the server in opencode.
- Requests without `jsonrpc: "2.0"` or with a non string/number id now get -32600 instead of running. - `decide --images` gets the `[]` default the other repeatable options have, so the generated ai_decide schema takes an array of images. - An empty value for a required input, such as video start `output`, counts as missing, so no video is submitted without its operation file. - A failed run with no output reports its exit code or signal. Adds a stdio test that a hung child is killed at its timeout; it fails if the process-group kill is removed.
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Tools keep the binary-name prefix (ai_*) and the opencode examples use the server key "ai", matching agent-browser's agent_browser_* tools under the "agent-browser" key. Reverts the cli key change.
ctate
approved these changes
Oct 23, 2026
An agent could pass `--base-url` through extraArgs and point image, video, or models at its own endpoint. The child inherits the server's AI_GATEWAY_API_KEY, so the request carried the key as a bearer token. Gateway connection flags (--base-url, --headers, --team-id-or-slug) are now rejected in extraArgs in both the split and `=` forms, and were already hidden from typed inputs. The blocked flags come from each command's definition, so a new command with these options is covered. Operators set AI_GATEWAY_BASE_URL and AI_GATEWAY_TEAM_ID_OR_SLUG in the server environment instead. The models tool hint no longer suggests --base-url.
![vercel[bot]](https://avatars.githubusercontent.com/in/8329?s=60&v=4){kind=small}
Move setup to the top of the MCP page and README section. It says the server reads AI_GATEWAY_API_KEY from its environment and gives a copy-paste setup for Claude Code, Codex, opencode, and mcpServers clients, plus a table of the variables the server reads.
The README MCP section repeated the website page. It now shows the mcpServers config with the API key and links to the MCP docs for other clients, profiles, and tools. The docs page leads with the mcpServers config that most clients accept, then opencode, Claude Code, and Codex.
fx is the first setup example on the MCP page. Placeholders now read your-vercel-ai-gateway-api-key, and the README links to the MCP page for configuring any MCP client rather than naming a few.
An empty string skipped type checks for every input, so `extraArgs: ""` reached the connection-flag check as a string and threw, taking the whole server down. Only required inputs now treat "" as missing; other inputs are type-checked as usual. An unexpected error inside a request now returns a JSON-RPC -32603 error instead of exiting the server.
Many MCP clients pass only a few variables to the servers they start, so exporting the key in the shell does not reach the server. The MCP Inspector showed this: the key reached the server only through its -e flag.
MCP clients already put the server name in front of each tool, so the ai_ prefix showed up twice, as ai_ai_text in opencode. Tools are now text, image, decide, models, video_start, video_status, audio_speak, and audio_transcribe.
ctate
approved these changes
Oct 23, 2026
- Retain MCP tools and Gateway scope support. - Keep Gateway connection settings in the server environment.
ctate
approved these changes
Oct 29, 2026
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
aiby shelling out and guessing flags, with no typed contract to discover.ai mcpadds a stdio MCP server whose tools are generated from the CLI's own command definitions, so every call runs the sameaibinary and new flags show up automatically.Demo
opencode connected to
ai mcplists the tools and makes a typed decision withopenai/gpt-6-luna-decisions.demo.mp4